I
nd
o
ne
s
ia
n J
o
urna
l o
f
E
lect
rica
l En
g
ineering
a
nd
Co
m
pu
t
er
Science
Vo
l.
24
,
No
.
2
,
N
o
v
em
b
e
r
2
0
2
1
,
p
p
.
1
0
0
9
~
1
0
1
6
I
SS
N:
2
5
0
2
-
4
7
5
2
,
DOI
: 1
0
.
1
1
5
9
1
/ijeecs.v
24
.i
2
.
pp
1
0
0
9
-
1
0
1
6
1009
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs.ia
esco
r
e.
co
m
Dro
po
ut
,
a
ba
sic
a
nd
effectiv
e
regu
la
riza
tion
metho
d
for
a
dee
p
lea
rning
mo
del:
a
ca
se
study
B
ra
him
J
a
bir,
No
ureddin
e
F
a
lih
LIM
ATI
Lab
o
ra
to
r
y
,
S
u
lt
a
n
M
o
u
lay
S
li
m
a
n
e
Un
iv
e
rsity
,
Be
n
i
M
e
l
lal,
M
o
r
o
c
c
o
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
May
3
,
2
0
2
1
R
ev
is
ed
Sep
8
,
2
0
2
1
Acc
ep
ted
Sep
13
,
2
0
2
1
De
e
p
lea
rn
in
g
is
b
a
se
d
on
a
n
e
t
wo
rk
of
a
rti
ficia
l
n
e
u
r
o
n
s
in
s
p
ir
e
d
by
t
h
e
h
u
m
a
n
b
ra
in
.
T
h
is
n
e
two
rk
is
m
a
d
e
up
of
te
n
s
or
e
v
e
n
h
u
n
d
re
d
s
of
"
lay
e
rs"
of
n
e
u
ro
n
s.
Th
e
field
s
of
a
p
p
li
c
a
ti
o
n
of
d
e
e
p
lea
rn
i
n
g
a
re
i
n
d
e
e
d
m
u
lt
ip
le;
Ag
ricu
lt
u
re
is
one
of
th
o
se
field
s
in
w
h
ich
d
e
e
p
lea
rn
in
g
is
u
se
d
in
v
a
ri
o
u
s
a
g
ricu
lt
u
ra
l
p
ro
b
lem
s
(d
ise
a
se
d
e
tec
ti
o
n
,
p
e
st
d
e
tec
ti
o
n
,
a
n
d
we
e
d
id
e
n
ti
fica
ti
o
n
).
A
m
a
jo
r
p
ro
b
lem
with
d
e
e
p
lea
rn
i
n
g
is
h
o
w
to
c
re
a
te
a
m
o
d
e
l
th
a
t
wo
rk
s
we
ll
,
not
o
n
ly
on
t
h
e
lea
rn
in
g
se
t
b
u
t
a
lso
on
t
h
e
v
a
li
d
a
ti
o
n
se
t.
M
a
n
y
a
p
p
ro
a
c
h
e
s
u
se
d
in
n
e
u
ra
l
n
e
two
rk
s
a
re
e
x
p
li
c
it
l
y
d
e
sig
n
e
d
to
re
d
u
c
e
o
v
e
rfit
,
p
o
ss
ib
l
y
at
t
h
e
e
x
p
e
n
se
of
in
c
re
a
sin
g
v
a
li
d
a
ti
o
n
a
c
c
u
ra
c
y
a
n
d
trai
n
in
g
a
c
c
u
ra
c
y
.
In
th
is
p
a
p
e
r,
a
b
a
sic
t
e
c
h
n
iq
u
e
(
d
r
o
p
o
u
t
)
is
p
r
o
p
o
se
d
to
m
in
imiz
e
o
v
e
rfit
,
we
i
n
teg
ra
ted
it
in
t
o
a
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
m
o
d
e
l
to
c
las
sify
we
e
d
sp
e
c
ies
a
n
d
se
e
how
it
imp
a
c
ts
p
e
rfo
rm
a
n
c
e
,
a
c
o
m
p
lem
e
n
ta
ry
so
l
u
ti
o
n
(
e
x
p
o
n
e
n
ti
a
l
li
n
e
a
r
u
n
it
s
)
a
re
p
ro
p
o
se
d
to
o
p
t
imiz
e
th
e
o
b
tain
e
d
r
e
su
lt
s.
Th
e
re
su
lt
s
sh
o
we
d
t
h
a
t
t
h
e
se
p
r
o
p
o
se
d
so
l
u
ti
o
n
s
a
re
p
ra
c
ti
c
a
l
a
n
d
h
ig
h
l
y
a
c
c
u
ra
te,
e
n
a
b
li
n
g
us
to
a
d
o
p
t
th
e
m
in
d
e
e
p
lea
rn
in
g
m
o
d
e
ls.
K
ey
w
o
r
d
s
:
C
NN
Deep
lear
n
in
g
Dr
o
p
o
u
t
Ma
ch
in
e
lear
n
in
g
R
eg
u
lar
izatio
n
T
h
is
is
an
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r
th
e
CC
BY
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
B
r
ah
im
J
ab
ir
L
I
MA
T
I
L
ab
o
r
ato
r
y
Su
ltan
Mo
u
lay
Sli
m
an
e
Un
iv
e
r
s
ity
Mg
h
ila,
BP
592
B
en
i
Me
llal,
Mo
r
o
cc
o
E
m
ail:
ib
r
a.
jab
ir
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Ma
ch
in
e
lear
n
in
g
an
d
d
ee
p
le
ar
n
in
g
a
r
e
p
a
r
t
of
ar
tific
ial
in
t
ellig
en
ce
.
T
h
ese
ap
p
r
o
ac
h
es
b
o
th
r
esu
lt
in
em
p
o
we
r
in
g
co
m
p
u
ter
s
to
m
ak
e
in
tellig
en
t
d
ec
is
io
n
s
.
Ho
wev
er
,
d
ee
p
lear
n
in
g
is
a
s
u
b
c
ateg
o
r
y
of
m
ac
h
in
e
lear
n
in
g
b
ec
au
s
e
it
r
elies
on
u
n
atten
d
ed
lear
n
i
n
g
,
wh
ich
is
a
f
o
r
m
of
lear
n
in
g
b
ase
d
on
m
ath
em
atica
l
ap
p
r
o
ac
h
es
u
s
ed
to
m
o
d
el
d
ata
[
1
]
.
Deep
lear
n
in
g
ap
p
li
ca
tio
n
s
ar
e
u
s
ed
in
v
ar
io
u
s
s
ec
to
r
s
lik
e:
im
ag
e
r
ec
o
g
n
itio
n
,
au
t
o
m
atic
tr
an
s
l
atio
n
,
au
to
n
o
m
o
u
s
ca
r
,
m
ed
ical
d
iag
n
o
s
is
,
p
e
r
s
o
n
alize
d
r
ec
o
m
m
e
n
d
atio
n
s
,
au
to
m
atic
m
o
d
er
atio
n
of
s
o
cial
n
etwo
r
k
s
,
f
in
an
cial
p
r
ed
ictio
n
an
d
au
to
m
ate
d
tr
ad
in
g
,
id
en
tific
atio
n
of
d
ef
ec
tiv
e
p
ar
ts
.
T
h
e
f
ield
th
a
t
in
ter
ests
us
an
d
in
wh
ich
we
h
av
e
ex
p
e
r
im
en
ted
is
“a
g
r
icu
ltu
r
e”
[
2
]
,
[
3
]
,
in
d
ee
d
,
we
ca
n
u
s
e
it
f
o
r
th
e
d
etec
tio
n
of
wee
d
s
,
wate
r
m
a
n
a
g
em
en
t,
de
tectio
n
of
i
n
s
ec
ts
an
d
d
is
ea
s
es.
Deep
lear
n
in
g
is
a
n
etwo
r
k
th
at
is
m
ad
e
up
of
ten
s
or
ev
e
n
h
u
n
d
r
ed
s
of
“lay
er
s
”
of
n
eu
r
o
n
s
,
each
r
ec
eiv
in
g
an
d
in
ter
p
r
etin
g
in
f
o
r
m
atio
n
f
r
o
m
th
e
p
r
ev
i
o
u
s
lay
er
[
4
]
.
A
s
et
of
th
eo
r
ies
an
d
m
o
d
els
h
av
e
b
ee
n
b
r
o
u
g
h
t
in
ex
is
ten
ce
.
T
h
eir
u
n
if
ied
g
o
al
is
to
r
ea
c
h
h
i
g
h
er
ac
cu
r
ac
y
lev
els
th
at
can
be
ap
p
lied
to
s
o
lv
e
p
r
o
b
le
ms
in
s
ev
er
al
f
ield
s
,
a
g
r
icu
ltu
r
al,
in
d
u
s
tr
ial,
h
ea
lth
.
T
h
is
f
ield
alwa
y
s
r
em
ain
s
a
s
u
b
ject
of
r
esear
c
h
.
T
h
o
u
s
an
d
s
of
ex
p
e
r
im
en
ts
a
n
d
th
o
u
s
an
d
s
of
s
cien
tific
p
ap
er
s
h
av
e
b
ee
n
p
r
o
d
u
ce
d
in
d
ee
p
lear
n
in
g
a
n
d
m
ac
h
in
e
lear
n
in
g
in
all
a
p
p
li
ca
tio
n
f
ield
s
in
r
ec
en
t
y
ea
r
s
.
Ho
wev
er
,
t
h
e
r
esear
c
h
d
o
o
r
is
s
till
o
p
en
wh
er
e
s
cien
tis
t
s
ar
e
s
till
tr
y
in
g
to
r
ea
ch
b
etter
m
o
d
els
th
at
can
g
ain
an
im
p
o
r
tan
t
d
e
g
r
ee
of
lear
n
i
n
g
.
T
h
e
y
ar
e
t
r
y
in
g
to
d
is
co
v
er
all
th
e
s
ettin
g
s
th
at
af
f
ec
t
it,
s
tar
tin
g
f
r
o
m
d
ata
co
llectio
n
,
th
r
o
u
g
h
its
p
r
ep
a
r
atio
n
an
d
p
u
r
if
icatio
n
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci,
Vo
l.
24
,
No
.
2
,
No
v
em
b
er
2
0
2
1
:
1
0
0
9
-
1
0
1
6
1010
to
th
e
o
u
tp
u
t.
Pre
p
ar
in
g
a
d
ee
p
lear
n
in
g
m
o
d
el
th
at
can
r
ea
c
h
a
g
o
o
d
p
er
f
o
r
m
an
ce
is
a
tr
o
u
b
leso
m
e
is
s
u
e
[
5
]
.
A
m
o
d
el
with
to
o
little
ca
p
ac
ity
ca
n
n
o
t
g
ain
p
r
o
f
icien
c
y
with
th
e
p
r
o
b
lem
,
wh
ile
a
m
o
d
el
with
an
ex
ce
s
s
of
ca
p
ac
ity
can
lear
n
it
ex
ce
s
s
iv
ely
well
an
d
o
v
er
f
itti
n
g
[
6
]
.
S
e
v
e
r
a
l
a
p
p
r
o
a
c
h
e
s
a
p
p
e
a
r
e
d
to
r
e
d
u
c
e
t
h
e
g
e
n
e
r
a
l
i
z
a
t
i
o
n
o
v
e
r
f
i
t
t
u
r
n
a
b
o
u
t
u
s
i
n
g
a
l
a
r
g
e
r
m
o
d
e
l
t
h
a
t
m
a
y
be
i
m
p
o
r
t
a
n
t
to
u
t
i
l
i
z
e
r
e
g
u
l
a
t
i
o
n
d
u
r
i
n
g
t
r
a
i
n
i
n
g
t
h
a
t
k
e
e
p
s
t
h
e
w
e
i
g
h
t
of
t
h
e
m
o
d
e
l
l
i
t
t
l
e
.
T
h
e
s
e
s
t
r
a
t
e
g
i
e
s
d
i
m
i
n
i
s
h
o
v
e
r
f
i
t
t
i
n
g
,
y
e
t
t
h
e
y
can
l
i
k
e
w
i
s
e
p
r
o
m
p
t
f
a
s
t
e
r
m
o
d
e
l
o
p
t
i
m
i
z
a
t
i
o
n
a
n
d
b
e
t
t
e
r
p
e
r
f
o
r
m
a
n
c
e
.
A
m
o
n
g
t
h
e
s
e
t
e
c
h
n
i
q
u
e
s
[
7
]
,
we
can
d
i
s
t
i
n
g
u
i
s
h
e
m
p
i
r
i
c
a
l
m
e
t
h
o
d
s
(
d
r
o
p
o
u
t
,
d
r
o
p
c
o
n
n
e
c
t
,
s
t
o
c
h
a
s
t
i
c
p
o
o
l
i
n
g
)
a
n
d
e
x
p
l
i
c
i
t
s
t
r
a
t
e
g
i
e
s
(
w
e
i
g
h
t
d
e
g
r
a
d
a
t
i
o
n
,
n
e
t
w
o
r
k
s
i
z
e
a
d
j
u
s
t
m
e
n
t
)
[
8
]
.
S
t
i
l
l
,
t
h
e
s
e
m
e
a
n
s
m
u
s
t
be
s
u
b
j
e
c
t
to
c
o
n
t
r
o
l
s
a
n
d
r
u
l
e
s
t
h
a
t
we
w
i
l
l
t
r
y
to
d
i
s
c
u
s
s
.
In
t
h
i
s
a
r
t
i
c
l
e
,
we
w
i
l
l
t
r
a
i
n
a
d
e
e
p
l
e
a
r
n
i
n
g
m
o
d
e
l
on
our
p
r
e
p
a
r
e
d
d
a
t
a
s
e
t
,
d
i
s
c
u
s
s
o
v
e
r
f
i
t
t
i
ng
in
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
(
C
N
N
)
t
r
a
i
n
i
n
g
,
p
r
o
p
o
s
e
a
r
e
g
u
l
a
r
i
z
a
t
i
o
n
s
o
l
u
t
i
o
n
,
a
n
d
c
o
n
c
l
u
d
e
w
i
t
h
a
c
o
m
p
l
e
m
e
n
t
a
r
y
s
o
l
u
t
i
o
n
,
w
h
i
c
h
m
a
k
e
s
it
p
o
s
s
i
b
l
e
to
r
e
i
n
f
o
r
c
e
a
n
d
o
p
t
i
m
i
z
e
t
h
e
r
e
s
u
l
t
s
o
b
t
a
i
n
e
d
.
2.
M
AT
E
R
I
E
L
AND
M
E
T
H
O
DS
In
th
is
s
tu
d
y
,
we
u
s
ed
a
s
et
of
tech
n
i
q
u
es
an
d
m
eth
o
d
s
t
h
at
h
av
e
a
d
ir
ec
t
r
elatio
n
s
h
ip
with
d
ee
p
lear
n
in
g
,
wh
er
e
we
u
s
ed
t
h
e
lib
r
ar
ies
of
Ker
as
an
d
T
en
s
o
r
f
lo
w
f
o
r
Py
th
o
n
in
o
r
d
er
to
b
u
ild
an
d
tr
ai
n
our
m
o
d
el,
we
u
s
ed
T
e
n
s
o
r
b
o
ar
d
to
ev
alu
ate
its
p
er
f
o
r
m
an
ce
,
we
u
s
ed
a
d
ataset
to
co
n
d
u
ct
th
e
tr
ain
in
g
[
9
]
.
We
will
d
is
cu
s
s
th
e
r
eg
u
lar
izatio
n
m
eth
o
d
s
u
s
ed
d
u
r
in
g
th
is
s
tu
d
y
to
p
r
e
v
en
t
tr
ai
n
in
g
o
v
er
f
it,
an
d
we
will
tr
y
to
s
tan
d
at
each
one
d
u
r
in
g
th
e
f
o
llo
win
g
p
ar
ag
r
ap
h
s
.
2
.
1
.
T
o
o
ls
a
nd
lib
ra
ries
We
h
a
v
e
u
s
e
d
T
e
n
s
o
r
f
l
o
w
a
n
d
K
e
r
a
s
l
i
b
r
a
r
i
e
s
f
o
r
our
e
x
p
e
r
i
m
e
n
t
a
t
i
o
n
.
T
e
n
s
o
r
f
l
o
w
is
an
o
p
e
n
-
s
o
u
r
c
e
p
l
a
t
f
o
r
m
f
o
r
a
r
t
i
f
i
c
i
a
l
i
n
t
e
l
l
i
g
e
n
c
e
(
AI
)
.
It
h
a
s
an
e
x
t
e
n
s
i
v
e
,
a
d
a
p
t
a
b
l
e
e
n
v
i
r
o
n
m
e
n
t
of
d
e
v
i
c
e
s
a
n
d
l
i
b
r
a
r
i
e
s
[
1
0
]
.
T
h
i
s
t
o
o
l
h
e
l
p
e
d
us
to
b
u
i
l
d
our
C
N
N
m
o
d
e
l
.
K
e
r
a
s
is
t
he
m
o
s
t
w
i
d
e
l
y
u
s
e
d
P
y
t
h
o
n
t
o
o
l
in
t
h
e
w
o
r
l
d
f
o
r
d
e
e
p
l
e
a
r
n
i
n
g
.
T
h
i
s
o
p
e
n
-
s
o
u
r
c
e
l
i
b
r
a
r
y
,
c
r
e
a
t
e
d
by
F
r
a
n
ç
o
i
s
C
h
o
l
l
e
t
,
e
a
s
i
l
y
a
n
d
q
u
i
c
k
l
y
c
r
e
a
t
e
s
n
e
u
r
a
l
n
e
t
w
o
r
k
s
b
a
s
e
d
on
t
h
e
m
a
i
n
f
r
a
m
e
w
o
r
k
s
(
T
e
n
s
o
r
f
l
o
w
,
P
y
t
o
r
c
h
a
n
d
M
X
N
E
T
)
[
1
1
]
.
E
v
a
l
u
a
t
i
o
n
of
t
h
e
m
o
d
e
l
is
d
o
n
e
on
t
h
e
T
e
n
s
o
r
b
o
a
r
d
t
o
o
l
,
w
h
i
c
h
is
a
t
o
o
l
to
v
i
s
u
a
l
i
z
e
t
h
e
o
b
t
a
i
n
e
d
r
e
s
u
l
t
s
.
It
a
l
s
o
a
l
l
o
w
s
us
to
v
i
e
w
t
h
e
s
t
r
u
c
t
u
r
e
of
t
h
e
m
o
d
e
l
as
a
g
r
a
p
h
[
1
2
]
.
It
is
l
a
u
n
c
h
e
d
f
r
o
m
a
t
e
r
m
i
n
a
l
c
o
m
m
a
n
d
(
t
e
n
s
o
r
b
o
a
r
d
--
l
o
g
d
i
r
=
f
o
l
d
e
r
/
)
.
2
.
2
.
D
a
t
a
s
et
T
h
e
d
ataset
is
one
of
th
e
s
i
g
n
if
ican
t
f
ac
to
r
s
af
f
ec
tin
g
th
e
q
u
ality
of
th
e
tr
ain
in
g
m
o
d
els,
an
d
wh
en
ev
er
we
h
av
e
a
lar
g
e
d
ataset
well
p
r
ep
ar
ed
,
we
h
av
e
h
ig
h
tr
ain
in
g
ac
c
u
r
ac
y
[
1
3
]
.
T
h
e
d
ataset
u
s
ed
d
u
r
in
g
th
is
ex
p
e
r
im
en
t
is
a
d
a
taset
th
at
we
p
r
ev
io
u
s
ly
u
s
ed
d
u
r
in
g
a
s
tu
d
y
on
o
b
ject
d
ete
ctio
n
th
at
id
e
n
tifie
s
wee
d
s
u
s
in
g
C
NN
[
1
4
]
,
[
1
5
]
.
It
co
n
tain
s
1
9
3
2
im
ag
es
of
f
o
u
r
t
y
p
es
of
wee
d
s
as
s
h
o
wn
in
F
ig
u
r
e
1;
th
ese
im
ag
es
ar
e
co
llected
in
wh
ea
t
f
ield
s
with
a
p
r
o
f
ess
io
n
al
ca
n
o
n
ca
m
er
a
E
OS
7
0
0
D.
T
h
at
o
f
f
er
lar
g
e
d
atasets
in
d
if
f
er
en
t
d
o
m
ain
s
in
ten
d
ed
f
o
r
th
e
d
ata
s
cien
ce
co
m
m
u
n
i
ty
u
s
ed
to
ac
h
iev
e
d
ata
s
cien
ce
g
o
als.
We
m
ad
e
m
o
d
if
icatio
n
s
to
th
ese
im
ag
es
(
r
esize,
co
n
tr
ast
an
d
tile
)
.
We
also
ad
d
ed
d
ata
-
au
g
m
en
tatio
n
tech
n
iq
u
es
(
cr
o
p
,
r
o
tatio
n
an
d
f
lip
)
to
p
r
o
d
u
ce
more
tr
ain
in
g
im
a
g
es
th
at
can
r
aise
th
e
m
o
d
el'
s
ac
cu
r
ac
y
[
1
6
]
.
We
now
h
a
v
e
ar
o
u
n
d
3
0
0
0
im
a
g
es
b
elo
n
g
in
g
to
f
o
u
r
d
i
f
f
er
en
t
class
es.
Fig
u
r
e
1
.
T
h
e
d
ataset
s
am
p
les
2
.
3
.
Re
g
ula
riza
t
io
n
T
h
e
FC
lay
er
s
(
f
u
lly
co
n
n
ec
t
ed
)
tak
e
up
m
o
s
t
of
C
NN'
s
m
em
o
r
y
.
Mo
r
eo
v
er
,
th
e
co
n
c
ep
t
of
FC
cr
ea
tes
an
e
x
p
o
n
en
tial
m
em
o
r
y
p
r
o
b
lem
ca
lled
"o
v
e
r
f
itti
n
g
"
(
"o
v
er
-
c
o
n
n
ec
tio
n
"
lead
i
n
g
to
o
v
e
r
-
lear
n
in
g
)
,
s
lo
win
g
d
o
wn
th
e
p
r
o
ce
s
s
in
g
of
in
f
o
r
m
atio
n
,
wh
ic
h
p
u
s
h
es
th
e
m
o
d
el
to
f
it
t
o
o
well
to
t
h
e
tr
ain
in
g
s
et,
b
u
t
d
if
f
icu
lt
to
g
en
e
r
alize
to
n
ew
ex
am
p
les
th
at
wer
e
not
in
th
e
tr
ain
in
g
s
et.
In
an
o
th
er
way
,
th
e
m
o
d
el
p
e
r
ce
iv
es
s
p
ec
if
ic
im
ag
es
in
th
e
tr
ain
in
g
s
et
r
ath
er
th
an
g
e
n
er
al
p
atte
r
n
s
an
d
th
e
tr
ain
in
g
ac
cu
r
ac
y
will
be
h
ig
h
er
th
an
th
e
v
alid
atio
n
[
1
7
]
.
R
eg
u
lar
iza
tio
n
is
a
p
r
o
ce
s
s
aim
ed
at
av
o
i
d
in
g
t
h
is
p
r
o
b
lem
of
o
v
er
-
lear
n
in
g
,
wh
ich
r
esu
lts
f
r
o
m
an
e
x
ce
s
s
iv
e
ad
ap
tatio
n
of
th
e
m
o
d
el
to
t
h
e
tr
ain
in
g
d
ata
[
1
8
]
.
T
h
er
e
ar
e
r
e
g
u
lar
izatio
n
m
eth
o
d
s
to
r
ed
u
ce
th
e
o
v
e
r
f
itti
n
g
class
if
ied
b
etwe
en
e
m
p
ir
ical
an
d
e
x
p
licit
m
eth
o
d
s
as
s
h
o
wn
in
T
a
b
le
1
.
We
f
o
cu
s
on
d
r
o
p
o
u
t
in
th
is
p
a
p
er
an
d
tr
y
with
a
s
u
itab
le
s
o
lu
tio
n
to
o
p
tim
ize
th
e
m
o
d
el.
T
h
e
d
r
o
p
o
u
t
is
u
s
ed
to
r
an
d
o
m
ly
"tu
r
n
o
f
f
"
o
r
"ig
n
o
r
e"
n
eu
r
o
n
s
(
with
a
p
r
e
d
ef
in
ed
p
r
o
b
a
b
ilit
y
,
o
f
ten
ev
er
y
o
th
er
n
e
u
r
o
n
)
as
well
as
p
er
ip
h
er
al
n
eu
r
o
n
s
.
W
h
en
n
e
u
r
o
n
s
a
r
e
r
a
n
d
o
m
ly
"t
u
r
n
ed
o
f
f
"
f
r
o
m
t
h
e
n
etwo
r
k
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
Dro
p
o
u
t,
a
b
a
s
ic
a
n
d
effec
tive
r
eg
u
la
r
iz
a
tio
n
meth
o
d
fo
r
a
d
ee
p
lea
r
n
in
g
mo
d
el:
A
c
a
s
e
s
tu
d
y
(
B
r
a
h
im
J
a
b
ir
)
1011
d
u
r
in
g
lea
r
n
in
g
,
th
e
o
th
e
r
n
e
u
r
o
n
s
will
h
av
e
to
s
tep
in
a
n
d
h
an
d
le
th
e
r
ep
r
esen
tatio
n
r
eq
u
ir
ed
t
o
m
ak
e
p
r
ed
ictio
n
s
f
o
r
th
e
m
is
s
in
g
n
e
u
r
o
n
s
[
2
5
]
.
T
h
u
s
,
with
f
ewe
r
n
eu
r
o
n
s
,
th
e
n
etwo
r
k
is
m
o
r
e
r
esp
o
n
s
iv
e
a
n
d
ca
n
lear
n
f
aster
.
At
th
e
en
d
o
f
th
e
lear
n
in
g
s
ess
io
n
,
th
e
"tu
r
n
e
d
o
f
f
"
n
eu
r
o
n
s
ar
e
"tu
r
n
ed
b
ac
k
o
n
"
(
with
th
eir
o
r
ig
in
al
weig
h
ts
)
.
T
h
e
clo
s
er
th
e
f
u
lly
-
co
n
n
ec
ted
lay
er
is
to
th
e
s
o
u
r
ce
im
ag
e,
th
e
f
ew
er
n
eu
r
o
n
s
will
b
e
ex
tin
g
u
is
h
ed
as sh
o
wn
in
Fig
u
r
e
2
.
Du
r
in
g
th
e
lear
n
in
g
p
h
ase,
f
o
r
ea
ch
iter
atio
n
,
a
n
eu
r
o
n
is
k
ep
t
with
a
p
r
o
b
ab
ilit
y
p
.
Oth
e
r
wis
e,
it
i
s
d
elete
d
.
Du
r
in
g
th
e
test
p
h
ase
,
all
n
eu
r
o
n
s
ar
e
k
e
p
t,
s
o
we
wan
t
th
e
n
eu
r
o
n
s
'
o
u
tp
u
ts
at
th
e
tim
e
o
f
test
in
g
to
b
e
th
e
s
am
e
as
th
eir
o
u
t
p
u
ts
at
th
e
tim
e
o
f
lear
n
in
g
.
Fo
r
e
x
am
p
le,
i
n
th
e
ca
s
e
wh
er
e
th
e
v
alu
e
o
f
d
r
o
p
o
u
t
-
0
.
2
5
,
n
eu
r
o
n
s
m
u
s
t
r
ed
u
ce
t
h
e
ir
p
r
o
d
u
ctio
n
(
o
u
tp
u
t)
b
y
2
5
%
at
th
e
tim
e
o
f
th
e
test
to
h
av
e
th
e
s
am
e
o
u
t
p
u
t
as
d
u
r
in
g
th
e
tr
ain
in
g
.
T
ab
le
1
.
R
eg
u
lar
izatio
n
m
eth
o
d
s
M
e
t
h
o
d
Em
p
i
r
i
c
M
e
t
h
o
d
E
x
p
l
i
c
i
t
D
r
o
p
o
u
t
[
1
9
]
W
e
i
g
h
t
d
e
g
r
a
d
a
t
i
o
n
[
2
2
]
D
r
o
p
C
o
n
n
e
c
t
[
2
0
]
A
d
j
u
st
t
h
e
n
e
t
w
o
r
k
s
i
z
e
[
2
3
]
P
o
o
l
i
n
g
st
o
c
h
a
st
i
q
u
e
[
2
1
]
B
a
t
c
h
n
o
r
ma
l
i
z
a
t
i
o
n
[
2
4
]
(
a)
(
b
)
Fig
u
r
e
2
.
Ap
p
ly
in
g
d
r
o
p
o
u
t
to
p
r
ev
en
t
n
eu
r
al
n
etwo
r
k
s
f
r
o
m
o
v
er
f
itti
n
g
[
2
6
]
; (
a)
s
tan
d
a
r
d
n
eu
tr
al
n
et
a
n
d
(
b
)
a
f
ter
ap
p
ly
i
n
g
d
r
o
p
o
u
t
2
.
4
.
M
o
del
a
rc
hite
ct
ure
Ou
r
g
o
al
d
u
r
i
n
g
th
is
s
tu
d
y
is
to
d
is
co
v
er
th
e
r
eg
u
lar
izatio
n
m
eth
o
d
s
an
d
h
o
w
d
r
o
p
o
u
t
is
an
im
p
o
r
tan
t
f
ac
to
r
th
at
elim
in
ates
o
v
er
f
itti
n
g
;
f
o
r
th
is
,
we
im
p
lem
en
te
d
a
s
im
p
le
C
NN
m
o
d
el
an
d
d
ef
in
ed
th
eir
lay
er
s
an
d
h
y
p
er
p
ar
am
eter
s
in
o
r
d
er
to
t
r
ain
th
em
on
o
u
r
p
r
e
p
ar
ed
d
a
taset
[
2
7
]
.
T
h
e
a
r
ch
itectu
r
e
u
s
ed
in
th
is
s
tu
d
y
is
s
tr
u
ctu
r
ed
:
C
o
n
v
→
Po
o
l
→
C
o
n
v
→
FC
→
Ou
tp
u
t
as
s
h
o
wn
in
Fig
u
r
e
3
.
First,
we
d
id
not
in
teg
er
d
r
o
p
o
u
t
to
an
aly
ze
th
e
r
esu
lts
b
ef
o
r
e
an
d
af
ter
d
r
o
p
o
u
t
.
T
h
e
m
o
d
el
is
r
u
n
on
20
e
p
o
ch
s
a
n
d
g
a
v
e
us
r
esu
lts
th
at
will
be
d
is
cu
s
s
ed
in
th
e
n
ex
t
s
ec
tio
n
.
Fig
u
r
e
3
.
Mo
d
el
ar
ch
itectu
r
e
b
ef
o
r
e
d
r
o
p
o
u
t
m
eth
o
d
3.
RE
SU
L
T
AND
DI
SCUS
SI
O
N
We
r
u
n
th
e
m
o
d
el
on
o
u
r
d
ata
s
et
in
20
s
tep
s
u
s
in
g
Py
th
o
n
.
T
h
e
r
esu
lts
o
b
tain
e
d
ar
e
v
is
u
al
ized
on
th
e
T
en
s
o
r
b
o
ar
d
to
o
l.
T
h
e
v
alid
atio
n
of
th
e
m
o
d
el
is
to
ev
alu
ate
th
e
ca
p
ac
ity
of
th
e
tr
ai
n
ed
m
o
d
el
to
g
en
er
alize
to
n
ew
s
am
p
les.
T
h
e
ac
cu
r
ac
y
of
th
e
tr
ain
in
g
was
also
u
s
ed
as
an
esti
m
ato
r
f
o
r
r
an
k
in
g
t
h
e
m
o
d
el
in
a
v
ar
iab
le
s
elec
tio
n
ap
p
r
o
ac
h
.
T
he
F
ig
u
r
e
4
s
h
o
ws
th
e
r
esu
lts
of
th
e
t
r
a
in
in
g
,
r
e
p
r
esen
ted
by
th
e
o
r
an
g
e
an
d
b
lu
e
cu
r
v
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci,
Vo
l.
24
,
No
.
2
,
No
v
em
b
er
2
0
2
1
:
1
0
0
9
-
1
0
1
6
1012
T
h
e
g
a
p
b
etwe
en
th
e
v
alid
at
io
n
ac
cu
r
ac
y
(
o
r
a
n
g
e
c
u
r
v
e
)
an
d
th
e
tr
ain
in
g
ac
c
u
r
ac
y
(
b
l
u
e
cu
r
v
e)
d
em
o
n
s
tr
at
es
th
e
q
u
an
tity
of
o
v
er
f
itti
n
g
as
s
h
o
wn
in
Fig
u
r
e
4
.
T
wo
p
o
ten
tial
ca
s
es
h
av
e
ap
p
ea
r
ed
in
th
e
ch
ar
t.
T
h
e
o
r
a
n
g
e
cu
r
v
e
(
v
alid
atio
n
ac
cu
r
ac
y
)
s
h
o
ws
ex
ce
p
tio
n
all
y
little
ap
p
r
o
v
al
ac
cu
r
ac
y
co
m
p
ar
ed
to
t
h
e
tr
ain
in
g
ac
cu
r
ac
y
.
Sh
o
win
g
h
ig
h
o
v
er
f
itti
n
g
im
p
lies
we
n
ee
d
to
r
eg
u
lar
ize
th
e
m
o
d
el
with
tech
n
iq
u
es
(
d
r
o
p
o
u
t
f
o
r
our
ca
s
e)
an
d
g
ath
er
m
o
r
e
in
f
o
r
m
atio
n
.
T
h
e
s
ec
o
n
d
c
o
n
ce
iv
ab
l
e
ca
s
e
is
wh
en
th
e
v
alid
atio
n
ac
cu
r
ac
y
tr
ac
k
s
th
e
tr
ain
in
g
ac
c
u
r
ac
y
g
e
n
u
in
ely
w
ell
[
2
8
]
.
T
h
is
ca
s
e
d
e
m
o
n
s
tr
at
es
th
at
th
e
m
o
d
el
ca
p
ac
ity
is
n
o
t
s
u
f
f
icien
tly
h
ig
h
.
It
got
us
th
in
k
in
g
ab
o
u
t
m
a
k
in
g
th
e
m
o
d
el
lar
g
er
by
ex
p
a
n
d
in
g
th
e
n
u
m
b
er
of
p
ar
am
et
er
s
.
T
h
e
f
o
llo
win
g
s
ec
tio
n
s
h
o
ws
us
th
e
s
o
lu
tio
n
s
p
r
o
p
o
s
ed
to
elim
in
ate
o
v
er
f
itt
in
g
an
d
ac
h
iev
e
b
etter
p
er
f
o
r
m
an
ce
.
3
.
1
.
Dr
o
po
ut
s
o
lutio
n
T
h
e
f
o
llo
win
g
m
o
d
el
is
th
e
s
am
e
ar
ch
itectu
r
e
but
r
e
g
u
lar
ize
d
with
th
e
tech
n
iq
u
e
of
d
r
o
p
o
u
t
.
Dr
o
p
o
u
t
0
.
5
m
ea
n
s
th
at
each
o
u
tp
u
t
n
e
u
r
o
n
f
r
o
m
th
e
f
u
lly
-
co
n
n
ec
ted
lay
er
h
as
a
50%
ch
a
n
ce
of
b
ei
n
g
k
e
p
t.
Mo
r
e
o
v
er
,
an
o
th
er
d
r
o
p
o
u
t
lay
er
at
th
e
ex
it
o
f
th
e
f
ir
s
t
d
en
s
e
with
th
e
p
r
o
b
a
b
ilit
y
of
2
5
%.
Dr
o
p
o
u
t
0
.
7
5
was
also
p
lace
d
af
ter
th
e
p
o
o
lin
g
p
h
ase
as
s
h
o
wn
in
F
ig
u
r
e
5.
T
h
is
tech
n
iq
u
e
r
an
d
o
m
ly
d
ea
ctiv
ates
n
e
u
r
o
n
s
in
o
u
r
n
etwo
r
k
so
th
at
it
is
r
ed
u
n
d
a
n
t
an
d
ca
n
f
in
d
n
ew
way
s
to
s
o
lv
e
th
e
o
v
e
r
f
i
ttin
g
p
r
o
b
lem
.
Fro
m
T
en
s
o
r
b
o
ar
d
,
we
g
et
th
e
g
r
ap
h
o
f
tr
ain
i
n
g
ac
c
u
r
ac
y
a
n
d
v
alid
atio
n
ac
c
u
r
ac
y
,
af
ter
t
r
ain
in
g
o
f
o
u
r
m
o
d
if
ie
d
m
o
d
el
ab
o
v
e
e
q
u
ip
p
e
d
with
d
r
o
p
o
u
t
.
T
h
e
Fig
u
r
e
6
s
h
o
ws
th
e
ev
o
lu
tio
n
o
f
th
e
v
alid
atio
n
r
ep
r
esen
ted
b
y
th
e
b
lu
e
cu
r
v
e,
an
d
t
h
e
ev
o
lu
tio
n
o
f
th
e
tr
ain
i
n
g
r
e
p
r
esen
ted
b
y
th
e
o
r
a
n
g
e
cu
r
v
e,
t
h
ese
r
esu
lts
ar
e
ex
p
lain
e
d
in
th
e
n
e
x
t
p
a
r
t.
Fro
m
th
e
r
esu
lts
,
we
ca
n
im
m
ed
iately
s
ee
th
at
th
e
m
o
d
el
wit
h
d
r
o
p
o
u
t
ac
h
iev
es
a
g
o
o
d
p
er
f
o
r
m
an
ce
.
W
ith
m
o
d
el
2
,
th
e
v
alid
atio
n
c
u
r
v
e
(
o
r
an
g
e)
tr
ac
k
s
th
at
o
f
th
e
tr
ai
n
(
b
lu
e)
,
wh
ich
m
ea
n
s
o
v
er
f
itti
n
g
is
q
u
ick
ly
m
in
im
iz
ed
ev
en
with
f
ewe
r
n
eu
r
o
n
s
d
u
r
in
g
lear
n
in
g
(
th
e
ef
f
ec
t
o
f
d
r
o
p
o
u
t)
.
T
h
is
ef
f
ec
t
ca
n
b
e
s
ee
n
in
th
e
p
r
ec
is
io
n
w
h
er
e
th
e
p
er
ce
n
tag
e
o
f
g
o
o
d
cl
ass
if
icatio
n
r
ea
ch
ed
an
d
6
0
% a
f
t
er
6
ep
o
c
h
s
.
T
h
is
p
r
ec
is
io
n
co
n
tin
u
es
to
in
cr
ea
s
e
s
lo
wly
,
r
ea
ch
in
g
8
5
%
af
ter
2
0
ep
o
ch
s
.
W
e
ca
n
r
elea
s
e
th
a
t
th
e
ea
s
ie
s
t
way
to
lim
it
o
v
er
f
itti
n
g
is
to
in
tr
o
d
u
c
e
th
e
d
r
o
p
o
u
t
tech
n
iq
u
e.
T
h
e
l
o
ca
tio
n
o
f
t
h
is
tech
n
iq
u
e
(
d
r
o
p
o
u
t)
ca
n
in
f
lu
en
ce
th
e
r
esu
lts
,
ev
en
if
we
ca
n
b
e
ap
p
lied
f
o
r
ea
c
h
lay
er
o
f
th
e
n
etwo
r
k
o
r
af
ter
s
elec
ted
lay
er
s
,
s
o
we
h
a
v
e
to
tr
y
d
if
f
er
en
t
co
m
b
in
atio
n
s
to
g
et
th
e
b
est
r
esu
lts
.
Al
s
o
,
th
e
f
ix
ed
d
r
o
p
o
u
t
p
r
o
b
a
b
ilit
ies
d
ir
ec
tly
in
f
lu
en
ce
th
e
o
v
er
f
itti
n
g
.
T
h
e
g
a
p
wer
e
p
ar
tially
r
ed
u
ce
d
h
o
wev
er
,
it
n
e
ed
s
m
o
r
e
r
e
g
u
lar
izatio
n
a
n
d
n
ee
d
im
p
r
o
v
in
g
th
e
ab
ilit
y
o
f
th
e
m
o
d
el
to
g
en
er
alize
.
T
h
e
p
r
o
p
o
s
ed
s
o
lu
tio
n
f
o
r
th
ese
p
r
o
b
le
m
s
will
b
e
d
i
s
cu
s
s
ed
in
th
e
n
ex
t
s
ec
tio
n
.
Fig
u
r
e
4
.
T
r
ain
in
g
a
n
d
v
alid
atio
n
ac
cr
u
r
ac
y
Fig
u
r
e
5
.
Mo
d
el
ar
ch
itectu
r
e
a
f
ter
d
r
o
p
o
u
t
m
eth
o
d
Fig
u
r
e
6.
T
r
ain
in
g
a
n
d
v
alid
atio
n
ac
cr
u
r
ac
y
af
te
r
d
r
o
p
o
u
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
Dro
p
o
u
t,
a
b
a
s
ic
a
n
d
effec
tive
r
eg
u
la
r
iz
a
tio
n
meth
o
d
fo
r
a
d
ee
p
lea
r
n
in
g
mo
d
el:
A
c
a
s
e
s
tu
d
y
(
B
r
a
h
im
J
a
b
ir
)
1013
3
.
2
.
O
ptim
iza
t
i
o
n
T
h
e
d
r
o
p
o
u
t
tech
n
iq
u
e
r
esu
lts
ar
e
in
ter
esti
n
g
,
but
o
v
er
f
it
is
e
lim
in
ated
,
an
d
th
e
ac
cu
r
ac
y
n
ee
d
s
to
be
b
o
o
s
ted
ev
en
m
o
r
e.
Fo
r
th
is
,
we
will
p
r
o
p
o
s
e
an
o
p
tim
iz
atio
n
s
o
lu
tio
n
th
at
can
be
ac
co
m
p
an
ied
with
th
e
d
r
o
p
o
u
t
tech
n
iq
u
e.
In
t
h
is
p
ar
t,
we
will
d
is
cu
s
s
a
s
o
lu
tio
n
f
o
r
o
p
tim
izin
g
th
e
r
esu
lts
o
b
ta
in
ed
b
ased
on
th
e
ac
tiv
atio
n
f
u
n
ctio
n
.
C
u
r
r
e
n
tly
,
th
e
m
o
s
t
p
o
p
u
lar
ac
tiv
atio
n
f
u
n
ctio
n
f
o
r
n
e
u
r
al
n
etwo
r
k
s
is
(
R
eL
u
)
.
T
h
e
R
eL
u
(
r
ec
tifie
d
lin
ea
r
u
n
it)
ac
tiv
ati
o
n
f
u
n
ctio
n
is
th
e
id
en
tity
f
o
r
p
o
s
itiv
e
in
p
u
ts
an
d
ze
r
o
es
o
t
h
er
wis
e.
T
h
e
m
ain
ad
v
an
tag
e
of
R
eL
u
is
th
at
it
s
o
lv
es
th
e
v
an
is
h
in
g
g
r
ad
ien
t
p
r
o
b
lem
.
Ho
wev
er
,
t
he
F
ig
u
r
e
7
s
h
o
ws
th
at
th
e
R
eL
u
(
p
in
k
cu
r
v
e)
is
not
n
eg
ativ
e
an
d
th
er
ef
o
r
e
h
as
an
av
e
r
ag
e
ac
tiv
atio
n
g
r
ea
te
r
th
an
z
er
o
.
Un
its
th
at
h
av
e
an
av
e
r
ag
e
ac
ti
v
atio
n
o
th
er
th
an
ze
r
o
act
as
a
b
ias
f
o
r
th
e
n
ex
t
lay
er
.
If
th
ese
u
n
its
do
n
o
t
ca
n
ce
l
ea
c
h
o
th
er
o
u
t,
th
e
tr
ain
in
g
ca
u
s
es
a
b
ias
s
h
if
t
f
o
r
th
e
u
n
its
in
th
e
n
e
x
t
lay
er
.
E
x
p
o
n
e
n
tial
lin
ea
r
u
n
its
(
ELU
)
lik
e
R
eL
u
,
ad
d
r
ess
es
th
e
v
an
is
h
in
g
g
r
ad
i
en
t
p
r
o
b
lem
with
id
en
tity
f
o
r
p
o
s
itiv
e
in
p
u
ts
.
T
h
e
s
am
e
f
ig
u
r
e
in
d
icate
s
th
at,
co
m
p
ar
ed
to
R
eL
u
,
ELU
(
p
u
r
p
le
cu
r
v
e
)
im
p
r
o
v
es
lear
n
in
g
b
ec
au
s
e
it
h
as
n
eg
ativ
e
v
alu
es
th
at
allo
w
it
to
p
u
s
h
th
e
av
er
ag
e
u
n
it
ac
tiv
atio
n
s
clo
s
er
to
ze
r
o
,
wh
ich
s
p
ee
d
s
up
lear
n
in
g
an
d
lead
s
to
b
etter
p
r
ec
is
io
n
in
class
if
icatio
n
.
T
h
e
ELU
f
u
n
ctio
n
is
d
ef
in
ed
in
th
e
p
y
th
o
n
c
o
d
e
with
th
e
f
o
llo
win
g
lin
e
«
K
era
s
.
la
ye
r
s
.
E
LU(
a
lp
h
a
=1
.
0
)
».
As
p
er
th
e
(
1
)
,
T
h
e
ELU
f
u
n
ctio
n
with
0
<
α
is
,
(
)
=
{
>
0
(
e
xp
(
)
−
1
)
≤
0
(
1
)
We
will
in
tr
o
d
u
ce
ELU
in
o
u
r
n
etwo
r
k
a
n
d
k
ee
p
o
th
er
lay
er
s
an
d
th
eir
o
r
d
er
as
in
t
h
e
p
r
ev
i
o
u
s
m
o
d
el,
th
en
th
e
ar
c
h
itectu
r
e
s
h
o
ws
in
Fig
u
r
e
8
.
We
h
av
e
tr
a
in
ed
th
e
n
ew
m
o
d
el
on
20
e
p
o
ch
s
.
We
th
en
o
b
tain
th
e
r
esu
lts
of
F
ig
u
r
e
9.
We
w
ill
co
m
p
ar
e
th
e
R
eL
U
m
o
d
el
as
s
h
o
wn
in
F
ig
u
r
e
5
an
d
th
at
of
th
e
ELU
m
o
d
el
p
er
f
o
r
m
an
ce
s
as
s
h
o
wn
in
F
ig
u
r
e
8
a
n
d
s
ee
wh
at
ch
a
n
g
es
will
o
cc
u
r
.
Fig
u
r
e
7
.
R
elu
an
d
E
L
U
(
el
u
,
alp
h
a=
1
)
Fig
u
r
e
8
.
R
ep
lacin
g
th
e
ac
tiv
a
tio
n
f
u
n
ctio
n
R
eL
u
b
y
ELU
Fig
u
r
e
9
.
L
ea
r
n
in
g
ev
o
lu
tio
n
o
f
th
e
E
L
U
m
o
d
el
T
h
e
ab
o
v
e
g
r
ap
h
s
h
o
ws
th
e
ab
ilit
y
of
ELU
to
g
en
er
alize
wh
er
e
th
e
s
co
r
e
of
9
7
%
is
r
ea
ch
ed
af
ter
20
ep
o
ch
s
.
We
th
en
r
ea
lize
th
at
th
e
ELU
ac
tiv
atio
n
f
u
n
ct
io
n
ac
ce
ler
ates
lear
n
in
g
co
m
p
ar
ed
to
R
eL
u
an
d
im
p
r
o
v
es
th
e
n
etwo
r
k
'
s
ab
ilit
y
to
g
en
er
alize
.
ELU
k
ee
p
s
s
o
m
e
of
th
e
p
o
s
itiv
e
th
in
g
s
to
f
i
x
s
o
m
e
of
th
e
R
eL
U
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci,
Vo
l.
24
,
No
.
2
,
No
v
em
b
er
2
0
2
1
:
1
0
0
9
-
1
0
1
6
1014
f
u
n
ctio
n
p
r
o
b
lem
s
[
2
9
]
.
T
h
er
e
ar
e
s
ev
er
al
in
ter
p
r
etatio
n
s
of
t
h
ese
tech
n
iq
u
es,
an
d
th
e
r
e
ar
e
m
an
y
r
ea
s
o
n
s
wh
y
th
ey
ar
e
p
r
ac
tical
an
d
e
f
f
ec
tiv
e
s
o
lu
tio
n
s
:
-
T
h
e
d
r
o
p
o
u
t
ca
n
be
in
ter
p
r
ete
d
as
a
s
am
p
lin
g
of
a
n
e
u
r
al
n
etwo
r
k
in
th
e
co
m
p
lete
n
eu
r
al
n
etwo
r
k
an
d
u
p
d
ate
o
n
ly
th
e
p
ar
am
eter
s
of
th
e
s
am
p
led
n
etwo
r
k
.
T
h
e
d
if
f
er
en
ce
is
th
at
th
ese
s
u
b
n
ets
s
h
ar
e
s
ettin
g
s
.
-
Oth
er
s
s
ee
d
r
o
p
o
u
t
as
a
f
o
r
m
of
d
ata
au
g
m
e
n
tatio
n
by
ar
ti
f
icially
co
r
r
u
p
tin
g
t
h
e
en
tr
ies
in
each
lay
er
.
T
h
is
g
r
ea
tly
e
x
p
an
d
s
th
e
n
u
m
b
er
of
ex
a
m
p
les
th
e
m
o
d
el
will
u
s
e
wh
en
t
r
ain
in
g
to
h
elp
p
r
o
tect
ag
ain
s
t
o
v
er
f
itti
n
g
.
-
An
o
th
er
in
ter
p
r
etatio
n
is
th
at
th
is
is
a
f
o
r
m
of
b
ag
g
in
g
in
wh
ich
a
s
et
of
p
atter
n
s
is
o
n
ly
tr
ain
ed
on
a
s
m
all
s
u
b
s
et
of
d
ata.
-
E
L
U
ad
d
r
ess
es
th
e
v
an
is
h
in
g
g
r
ad
ien
t
p
r
o
b
lem
with
id
en
tity
f
o
r
p
o
s
itiv
e
in
p
u
ts
.
-
ELU
im
p
r
o
v
es
lear
n
in
g
b
ec
a
u
s
e
it
h
as
n
eg
ativ
e
v
alu
es
th
at
a
llo
w
it
to
p
u
s
h
th
e
u
n
it'
s
av
er
a
g
e
ac
tiv
atio
n
s
clo
s
er
to
ze
r
o
,
w
h
ich
s
p
ee
d
s
up
lear
n
in
g
an
d
lea
d
s
to
b
etter
p
r
ec
is
io
n
in
class
if
icatio
n
.
No
w
we
h
av
e
o
b
tain
e
d
o
u
r
f
in
al
m
o
d
el
wh
er
e
th
e
d
r
o
p
o
u
t
a
n
d
ELU
ar
e
two
ess
en
tial
p
ar
a
m
eter
s
th
at
ef
f
ec
tiv
ely
r
ed
u
ce
o
v
er
f
itti
n
g
an
d
in
cr
ea
s
e
p
er
f
o
r
m
an
ce
.
Sin
ce
we
p
lan
to
g
et
th
e
m
o
s
t
out
of
our
n
etwo
r
k
,
we
p
lan
to
let
th
e
lear
n
in
g
go
on
f
o
r
a
v
er
y
l
o
n
g
tim
e
an
d
b
ac
k
up
t
h
e
s
ettin
g
s
at
each
ep
o
ch
u
n
til
th
er
e
is
no
f
u
r
th
er
im
p
r
o
v
e
m
en
t
in
p
er
f
o
r
m
an
ce
ac
r
o
s
s
th
e
d
ata.
Fin
ally
,
we
can
s
ay
th
at
id
en
tify
in
g
o
v
er
f
itti
n
g
is
u
s
ef
u
l,
y
et
it
d
o
es
not
tac
k
le
th
e
is
s
u
e.
L
u
ck
ily
,
we
h
av
e
more
ch
o
ic
es
to
attem
p
t
in
ad
d
itio
n
to
d
r
o
p
o
u
t
an
d
ELU
a
n
d
th
e
o
th
er
r
eg
u
lar
izatio
n
m
eth
o
d
s
.
T
h
o
s
e
o
th
er
m
eth
o
d
s
also
can
lim
it
o
v
er
f
itti
n
g
an
d
in
cr
ea
s
e
th
e
ac
cu
r
ac
y
,
we
cite:
-
C
r
o
s
s
-
v
alid
atio
n
:
Utilize
th
e
i
n
itial
tr
ain
in
g
d
ata
to
p
r
o
d
u
ce
n
u
m
er
o
u
s
more
m
i
n
o
r
th
an
e
x
p
ec
ted
tr
ain
-
test
s
p
lit
s
.
Utilize
th
ese
p
ar
ts
to
tu
n
e
th
e
m
o
d
el
[
3
0
]
.
-
I
n
cr
ea
s
e
th
e
d
atab
ase
s
ize:
T
r
ain
th
e
m
o
d
el
with
f
u
r
th
er
d
ata
can
i
m
p
r
o
v
e
al
g
o
r
ith
m
s
to
id
en
tif
y
t
h
e
s
ig
n
al
b
etter
[
3
1
]
.
-
E
ar
ly
s
to
p
p
in
g
:
co
n
s
is
ts
of
s
to
p
p
in
g
t
h
e
tr
ain
in
g
s
tep
as
s
o
o
n
as
th
e
v
alid
atio
n
lo
s
s
r
ea
ch
e
s
a
p
latea
u
or
in
cr
ea
s
es
[
3
2
]
.
-
B
atch
n
o
r
m
aliza
tio
n
:
E
ac
h
l
ay
er
o
b
s
er
v
es
in
p
u
ts
p
r
o
d
u
c
ed
by
lay
e
r
s
p
r
ec
ed
in
g
it.
It
wo
u
ld
be
ad
v
an
tag
e
o
u
s
to
be
ce
n
ter
ed
a
n
d
r
ed
u
ce
d
in
p
u
ts
f
o
r
each
lay
er
[
3
3
]
.
-
Glo
b
al
av
er
ag
e
p
o
o
lin
g
(
GAP
)
was
p
r
o
p
o
s
ed
to
r
ep
lace
th
e
m
u
lti
-
lay
er
ed
p
er
ce
p
tr
o
n
p
a
r
t.
T
h
e
id
ea
is
to
g
en
er
ate
a
f
ea
tu
r
e
m
ap
f
o
r
ea
c
h
co
r
r
esp
o
n
d
i
n
g
ca
teg
o
r
y
.
I
n
s
t
ea
d
of
ad
d
in
g
a
p
er
ce
p
tr
o
n
a
f
t
er
th
e
f
ea
t
u
r
e
m
ap
s
,
we
tak
e
th
e
av
er
a
g
e
of
each
of
th
em
,
an
d
th
e
r
esu
lt
is
in
s
er
ted
in
to
th
e
So
f
tm
ax
f
u
n
ctio
n
.
An
ad
v
an
tag
e
of
GAP
o
v
er
p
er
ce
p
tr
o
n
lay
e
r
s
is
th
at
th
er
e
ar
e
no
p
ar
am
eter
s
to
be
tr
ain
ed
;
t
h
er
ef
o
r
e,
o
v
er
-
lear
n
in
g
is
av
o
i
d
ed
[
3
4
]
.
4.
CO
NCLU
SI
O
N
C
NN
is
a
m
u
lti
-
lay
er
ed
n
e
u
r
al
n
etwo
r
k
t
h
at
is
u
s
ed
in
p
atter
n
an
d
im
ag
e
r
ec
o
g
n
itio
n
p
r
o
b
lem
s
.
Ma
n
y
p
r
o
b
lem
s
en
c
o
u
n
te
r
ed
d
u
r
in
g
lear
n
in
g
af
f
ec
t
th
e
m
o
d
el
p
er
f
o
r
m
an
ce
an
d
n
e
g
ativ
e
ly
in
f
lu
en
ce
r
esu
lts
o
b
tain
ed
;
am
o
n
g
th
ese
p
r
o
b
le
m
s
,
we
f
in
d
o
v
er
f
itti
n
g
,
w
h
ich
h
ap
p
en
s
wh
e
n
a
m
o
d
el
lear
n
s
th
e
d
etail
an
d
n
o
is
e
in
th
e
tr
ain
in
g
d
ata.
In
our
wo
r
k
,
we
u
s
ed
a
C
NN
f
o
r
im
ag
e
s
class
if
icatio
n
,
an
d
we
tr
ain
e
d
it
on
a
d
ataset,
we
d
is
co
v
er
th
e
p
r
o
b
lem
of
o
v
e
r
f
itti
n
g
wh
e
n
tr
ain
i
n
g
it,
an
d
th
en
we
p
r
o
p
o
s
ed
d
r
o
p
o
u
t
as
a
r
eg
u
lar
izatio
n
tech
n
iq
u
e
th
at
p
r
e
v
en
ts
to
o
v
er
f
it
p
r
o
b
lem
.
T
h
e
r
esu
lts
s
h
o
wed
th
at
th
e
ch
o
ice
of
th
e
p
r
o
b
ab
ilit
y
of
d
r
o
p
o
u
t
an
d
its
p
lace
s
ig
n
if
ican
tly
in
f
l
u
en
ce
s
r
esu
lts
.
B
u
t
th
e
r
esu
lts
ar
e
not
s
u
f
f
icien
t
f
o
r
th
at
we
s
u
g
g
ested
ad
d
i
n
g
an
ELU
f
u
n
ctio
n
to
o
p
tim
ize
th
e
ac
cu
r
ac
y
.
I
n
d
ee
d
,
it
p
o
s
itiv
el
y
af
f
ec
ted
th
e
r
esu
lts
,
en
ab
lin
g
us
to
r
ed
u
ce
th
e
o
v
er
f
it
an
d
r
ea
ch
97%
ac
cu
r
a
cy
.
In
f
u
tu
r
e
wo
r
k
,
we
will
s
tu
d
y
th
e
o
th
er
tec
h
n
iq
u
es
th
at
ad
d
r
ess
o
v
er
f
itti
n
g
an
d
b
r
in
g
o
u
t
b
etter
r
esu
lts
.
T
h
ey
will
be
r
ec
o
r
d
e
d
an
d
l
o
ad
ed
in
t
o
an
in
tellig
en
t
r
asp
b
er
r
y
-
b
ased
s
y
s
tem
,
en
ab
lin
g
th
e
r
ea
l
-
tim
e
id
e
n
tific
atio
n
of
wee
d
s
in
th
e
a
g
r
icu
ltu
r
al
en
v
ir
o
n
m
e
n
t
an
d
allo
win
g
th
em
to
be
s
p
r
ay
ed
lo
ca
lly
.
RE
F
E
R
E
NC
E
S
[1
]
M.
Ak
o
u
r
,
H.
Alsg
h
a
ier
,
a
n
d
O.
Al
Qa
se
m
,
"
Th
e
e
ffe
c
ti
v
e
n
e
ss
of
u
sin
g
d
e
e
p
lea
rn
i
n
g
a
lg
o
r
it
h
m
s
in
p
re
d
icti
n
g
stu
d
e
n
ts
a
c
h
ie
v
e
m
e
n
ts,"
In
d
o
n
e
si
a
n
J
o
u
rn
a
l
of
El
e
c
trica
l
E
n
g
in
e
e
rin
g
a
n
d
Co
m
p
u
ter
S
c
ie
n
c
e
(IJ
EE
CS
)
,
v
o
l
.
1
9
,
no.
1,
p
p
.
3
8
7
-
3
9
3
,
2
0
2
0
,
d
o
i:
1
0
.
1
1
5
9
1
/i
jee
c
s.v
1
9
.
i
1
.
p
p
3
8
8
-
3
9
4
.
[2
]
B.
Ja
b
ir,
N.
F
a
li
h
,
a
n
d
K.
Ra
h
m
a
n
i,
"
Ac
c
u
ra
c
y
a
n
d
Eff
icie
n
c
y
Co
m
p
a
ris
o
n
of
Ob
jec
t
De
tec
ti
o
n
O
p
e
n
-
S
o
u
rc
e
M
o
d
e
ls,
"
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
of
O
n
li
n
e
&
Bi
o
me
d
ic
a
l
E
n
g
in
e
e
rin
g
,
v
o
l.
1
7
,
n
o
.
5,
p
p
.
165
-
1
8
4
,
2
0
2
1
,
d
o
i:
1
0
.
3
9
9
1
/i
j
o
e
.
v
1
7
i
0
5
.
2
1
8
3
3
.
[3
]
Z.
Ün
a
l,
"
S
m
a
rt
fa
rm
in
g
b
e
c
o
m
e
s
e
v
e
n
sm
a
rter
with
d
e
e
p
lea
r
n
in
g
-
A
b
ib
li
o
g
ra
p
h
ica
l
a
n
a
ly
sis,
"
IEE
E
Acc
e
ss
,
v
o
l.
8,
p
p
.
1
0
5
5
8
7
-
1
0
5
6
0
9
,
2
0
2
0
,
d
o
i:
1
0
.
1
1
0
9
/ACCES
S
.
2
0
2
0
.
3
0
0
0
1
7
5
.
[4
]
S.
Do
n
g
,
P.
Wan
g
,
a
n
d
K.
Ab
b
a
s,
"A
su
r
v
e
y
on
d
e
e
p
lea
rn
in
g
a
n
d
its
a
p
p
li
c
a
ti
o
n
s,
"
C
o
mp
u
ter
S
c
ien
c
e
Rev
iew
,
v
o
l.
4
0
,
p.
1
0
0
3
7
9
,
2
0
2
1
,
d
o
i:
1
0
.
1
0
1
6
/j
.
c
o
sre
v
.
2
0
2
1
.
1
0
0
3
7
9
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
Dro
p
o
u
t,
a
b
a
s
ic
a
n
d
effec
tive
r
eg
u
la
r
iz
a
tio
n
meth
o
d
fo
r
a
d
ee
p
lea
r
n
in
g
mo
d
el:
A
c
a
s
e
s
tu
d
y
(
B
r
a
h
im
J
a
b
ir
)
1015
[5
]
S.
Re
h
m
a
n
,
S.
Tu
,
O.
Re
h
m
a
n
,
Y.
Hu
a
n
g
,
C.
M
a
g
u
ra
wa
lag
e
,
a
n
d
C.
-
C.
Ch
a
n
g
,
“
Op
ti
m
iza
ti
o
n
of
CNN
th
r
o
u
g
h
No
v
e
l
Train
in
g
S
trate
g
y
fo
r
Visu
a
l
Clas
sifica
ti
o
n
P
ro
b
lem
s,”
En
tro
p
y
,
v
o
l.
2
0
,
n
o
.
4,
p.
2
9
0
,
a
v
r.
2
0
1
8
,
d
o
i:
1
0
.
3
3
9
0
/e2
0
0
4
0
2
9
0
.
[6
]
M.
H.
Jo
p
ri,
A.
R.
Ab
d
u
ll
a
h
,
J.
T
o
o
,
T.
S
u
ti
k
n
o
,
S.
Ni
k
o
lo
v
sk
i
,
a
n
d
M.
M
a
n
a
p
,
"
S
u
p
p
o
rt
-
v
e
c
to
r
m
a
c
h
in
e
a
n
d
Na
ï
v
e
Ba
y
e
s
b
a
se
d
d
iag
n
o
stic
a
n
a
ly
ti
c
of
h
a
rm
o
n
ic
so
u
rc
e
id
e
n
ti
fica
ti
o
n
,
"
In
d
o
n
e
si
a
n
J
o
u
rn
a
l
of
El
e
c
trica
l
En
g
in
e
e
rin
g
and
Co
m
p
u
ter
S
c
ie
n
c
e
(IJ
EE
CS
)
,
v
o
l.
2
0
,
no.
1,
p
p
.
1
-
8,
2
0
2
0
,
d
o
i:
1
0
.
1
1
5
9
1
/i
jee
c
s.v
2
0
.
i
1
.
p
p
1
-
8.
[7
]
O.
De
n
iz,
A.
P
e
d
ra
z
a
,
N.
Va
ll
e
z
,
J.
S
a
li
d
o
,
a
n
d
G.
Bu
e
n
o
,
“
Ro
b
u
stn
e
ss
to
a
d
v
e
rsa
rial
e
x
a
m
p
les
can
be
imp
ro
v
e
d
with
o
v
e
rfit
ti
n
g
,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
of
M
a
c
h
in
e
L
e
a
rn
i
n
g
a
n
d
Cy
b
e
rn
e
ti
c
s
,
v
o
l.
1
1
,
n
o
.
4,
pp.
9
3
5
‑9
4
4
,
2
0
2
0
,
d
o
i:
1
0
.
1
0
0
7
/s1
3
0
4
2
-
0
2
0
-
0
1
0
9
7
-
4.
[8
]
V.
Ko
d
k
i
n
a
n
d
A.
S.
An
i
k
in
,
"
Th
e
e
x
p
e
rime
n
tal
i
d
e
n
ti
f
ica
ti
o
n
m
e
th
o
d
of
t
h
e
d
y
n
a
m
ic
e
fficie
n
c
y
f
o
r
fre
q
u
e
n
c
y
re
g
u
latio
n
a
l
g
o
ri
th
m
s
of
AEDs
,
"
In
ter
n
a
t
io
n
a
l
J
o
u
r
n
a
l
of
P
o
we
r
E
lec
tro
n
ics
a
n
d
Dr
ive
S
y
ste
ms
(IJ
PE
DS
)
,
v
o
l.
1
2
,
no.
1,
p
p
.
59
-
6
6
,
2
0
2
1
,
d
o
i:
1
0
.
1
1
5
9
1
/i
jp
e
d
s.v
1
2
.
i1
.
p
p
5
9
-
6
6
.
[9
]
T.
Do
lec
k
,
D.
J.
Lem
a
y
,
R.
B.
B
a
sn
e
t
,
a
n
d
P.
Ba
z
e
lais,
“
P
re
d
icti
v
e
a
n
a
ly
ti
c
s
in
e
d
u
c
a
ti
o
n
:
a
c
o
m
p
a
riso
n
of
d
e
e
p
lea
rn
in
g
fra
m
e
wo
rk
s,”
E
d
u
c
a
t
i
o
n
a
n
d
I
n
f
o
rm
a
ti
o
n
T
e
c
h
n
o
lo
g
ies
,
v
o
l.
2
5
,
n
o
.
3,
p
p
.
1
9
5
1
‑
1
9
6
3
,
2
0
2
0
,
d
o
i:
1
0
.
1
0
0
7
/s1
0
6
3
9
-
0
1
9
-
1
0
0
6
8
-
4.
[1
0
]
B.
P
a
n
g
,
E.
Nij
k
a
m
p
a
n
d
Y.
N.
Wu
,
“
De
e
p
lea
rn
in
g
wi
th
ten
so
rflo
w:
a
re
v
iew
,
”
J
o
u
rn
a
l
of
E
d
u
c
a
ti
o
n
a
l
a
n
d
Beh
a
v
io
r
a
l
S
ta
ti
st
ics
,
v
o
l.
45,
n
o
.
2,
p
p
.
2
2
7
-
2
4
8
,
2
0
2
0
,
1
0
.
3
1
0
2
/1
0
7
6
9
9
8
6
1
9
8
7
2
7
6
1
.
[1
1
]
A.
F
a
rk
a
s,
G.
Ke
r
tes
z
a
n
d
R.
Lo
v
a
s,
“
P
a
ra
ll
e
l
a
n
d
Distr
ib
u
ted
Trai
n
in
g
of
De
e
p
Ne
u
ra
l
Ne
t
wo
rk
s:
A
b
rief
o
v
e
rv
iew
,
”
in
2
0
2
0
IEE
E
2
4
t
h
I
n
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
on
In
tel
li
g
e
n
t
E
n
g
i
n
e
e
rin
g
S
y
ste
ms
(INE
S
)
,
Re
y
k
ja
v
ík
,
Ic
e
lan
d
,
Ju
l
y
2
0
2
0
,
p
p
.
1
6
5
‑1
7
0
,
d
o
i:
1
0
.
1
1
0
9
/INES
4
9
3
0
2
.
2
0
2
0
.
9
1
4
7
1
2
3
.
[1
2
]
D.
C.
Vo
g
e
lsa
n
g
a
n
d
B.
J.
Eri
c
k
so
n
,
“
M
a
g
icia
n
’s
C
o
rn
e
r:
6.
Ten
s
o
rF
lo
w
a
n
d
Ten
s
o
rBo
a
rd
,
”
R
a
d
i
o
lo
g
y
:
Arti
fi
c
i
a
l
In
telli
g
e
n
c
e
,
v
o
l.
2,
no.
3,
p.
e
2
0
0
0
1
2
,
M
a
y
2
0
2
0
,
d
o
i
:
1
0
.
1
1
4
8
/r
y
a
i
.
2
0
2
0
2
0
0
0
1
2
.
[1
3
]
A.
A.
Oju
g
o
a
n
d
R.
E.
Y
o
ro
,
"
F
o
rg
i
n
g
a
d
e
e
p
lea
rn
in
g
n
e
u
ra
l
n
e
two
rk
in
tr
u
sio
n
d
e
tec
ti
o
n
fra
m
e
wo
rk
to
c
u
rb
th
e
d
istri
b
u
ted
d
e
n
ial
of
se
rv
ice
a
tt
a
c
k
,
"
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
of
El
e
c
trica
l
a
n
d
C
o
mp
u
ter
E
n
g
in
e
e
rin
g
(IJ
ECE
)
,
v
o
l.
1
1
,
n
o
.
2,
p
p
.
1
4
9
8
-
1
5
0
9
,
2
0
2
1
,
d
o
i:
1
0
.
1
1
5
9
1
/
ij
e
c
e
.
v
1
1
i2
.
p
p
1
4
9
8
-
1
5
0
9
.
[1
4
]
B.
Ja
b
ir
a
n
d
N.
F
a
li
h
,
"
Di
g
it
a
l
a
g
ricu
l
tu
re
in
M
o
r
o
c
c
o
,
o
p
p
o
rtu
n
it
ies
a
n
d
c
h
a
ll
e
n
g
e
s,"
in
2
0
2
0
IEE
E
6
th
In
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
on
Op
ti
miza
ti
o
n
a
n
d
A
p
p
l
ica
ti
o
n
s
(ICOA)
.
IE
EE
,
2
0
2
0
.
p
p
.
1
-
5,
d
o
i:
1
0
.
1
1
0
9
/ICOA4
9
4
2
1
.
2
0
2
0
.
9
0
9
4
4
5
0
.
[1
5
]
B.
Ja
b
ir,
N.
S
.
F
a
li
h
,
a
n
d
A.
Tan
n
o
u
c
h
e
,
“A
S
trate
g
ic
An
a
l
y
ti
c
s
Us
in
g
Co
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
Ne
two
rk
s
f
o
r
Wee
d
Id
e
n
ti
fica
ti
o
n
in
S
u
g
a
r
Be
e
t
F
ield
s,"
AGRI
S
on
-
li
n
e
P
a
p
e
rs
in
Eco
n
o
mic
s
a
n
d
I
n
fo
rm
a
ti
c
s
,
v
o
l.
1
3
,
n
o
.
1,
pp.
49
-
5
7
,
2
0
2
1
,
d
o
i:
1
0
.
7
1
6
0
/a
o
l.
2
0
2
1
.
1
3
0
1
0
4
.
[1
6
]
R.
P
o
o
jary
,
R.
Ra
in
a
,
a
n
d
A.
Ku
m
a
r
M
o
n
d
a
l,
"
Eff
e
c
t
of
d
a
ta
-
a
u
g
m
e
n
tatio
n
on
fi
n
e
-
tu
n
e
d
CNN
m
o
d
e
l
p
e
rfo
rm
a
n
c
e
,
"
IAE
S
In
ter
n
a
ti
o
n
a
l
J
o
u
r
n
a
l
of
Arti
fi
c
ia
l
I
n
telli
g
e
n
c
e
(IJ
AI)
,
v
o
l.
1
0
,
n.
1,
p
p
.
84
-
92,
2
0
2
1
,
d
o
i:
1
0
.
1
1
5
9
1
/
ij
a
i.
v
1
0
.
i1
.
p
p
8
4
-
92.
[1
7
]
Q.
Xu
,
M.
Zh
a
n
g
,
Z.
Gu
,
a
n
d
G.
P
a
n
,
“
Ov
e
rfit
ti
n
g
re
m
e
d
y
by
s
p
a
rsify
in
g
re
g
u
lariz
a
ti
o
n
on
f
u
ll
y
-
c
o
n
n
e
c
ted
lay
e
rs
of
CNN
s,”
Ne
u
ro
c
o
mp
u
ti
n
g
,
v
o
l.
3
2
8
,
p
p
.
6
9
‑
7
4
,
2
0
1
9
,
d
o
i:
1
0
.
1
0
1
6
/j
.
n
e
u
c
o
m
.
2
0
1
8
.
0
3
.
0
8
0
.
[1
8
]
G.
F
a
ti
m
a
a
n
d
S.
S
a
e
e
d
,
"A
No
v
e
l
Weig
h
ted
E
n
se
m
b
le
M
e
th
o
d
to
Ov
e
rc
o
m
e
th
e
Im
p
a
c
t
of
Un
d
e
r
-
fi
tt
in
g
a
n
d
Ov
e
r
-
fit
ti
n
g
on
th
e
Clas
sifica
ti
o
n
Ac
c
u
ra
c
y
of
t
h
e
Im
b
a
lan
c
e
d
Da
ta
S
e
ts,"
P
a
k
ista
n
J
o
u
rn
a
l
o
f
S
ta
ti
st
ics
a
n
d
Op
e
ra
t
io
n
Res
e
a
rc
h
,
v
o
l.
17,
n
o
.
2,
pp.
4
8
3
‑
4
9
6
,
2
0
2
1
,
d
o
i:
1
0
.
1
8
1
8
7
/p
jso
r.
v
1
7
i2
.
3
6
4
0
.
[1
9
]
C.
G
a
rb
in
,
X.
Z
h
u
,
a
n
d
O.
M
a
rq
u
e
s,
"
Dro
p
o
u
t
v
s.
b
a
tch
n
o
rm
a
li
z
a
ti
o
n
:
an
e
m
p
iri
c
a
l
stu
d
y
of
th
e
ir
imp
a
c
t
to
d
e
e
p
lea
rn
in
g
,
"
M
u
lt
ime
d
ia
T
o
o
ls
a
n
d
Ap
p
li
c
a
t
io
n
s
,
v
o
l.
7
9
,
n
o
.
1
9
‑
2
0
,
p
p
.
1
2
7
7
7
‑
1
2
8
1
5
,
2
0
2
0
,
d
o
i:
1
0
.
1
0
0
7
/s1
1
0
4
2
-
0
1
9
-
0
8
4
5
3
-
9.
[2
0
]
Y.
S
a
k
a
i,
B.
U.
P
e
d
r
o
n
i
,
S.
J
o
sh
i,
S.
Ta
n
a
b
e
,
A.
A
k
in
i
n
,
a
n
d
G.
Ca
u
we
n
b
e
rg
h
s,
"
Dro
p
o
u
t
a
n
d
Dr
o
p
Co
n
n
e
c
t
f
o
r
Re
li
a
b
le
Ne
u
ro
m
o
r
p
h
ic
In
fe
re
n
c
e
Un
d
e
r
Co
m
m
u
n
ica
ti
o
n
C
o
n
stra
i
n
ts
in
Ne
two
r
k
Co
n
n
e
c
ti
v
i
ty
,
"
in
IEE
E
J
o
u
rn
a
l
on
Eme
rg
i
n
g
a
n
d
S
e
lec
ted
T
o
p
ics
in
Circ
u
it
s
a
n
d
S
y
ste
ms
,
v
o
l.
9,
no.
4,
p
p
.
6
5
8
-
6
6
7
,
De
c
.
2
0
1
9
,
d
o
i:
1
0
.
1
1
0
9
/JET
CAS.
2
0
1
9
.
2
9
5
2
6
4
2
.
[2
1
]
X.
Jia
n
g
,
M.
Lu
,
a
n
d
S.
H.
Wa
n
g
,
“
An
e
i
g
h
t
-
la
y
e
r
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
wit
h
st
o
c
h
a
sti
c
p
o
o
li
n
g
,
b
a
tch
n
o
rm
a
li
z
a
ti
o
n
a
n
d
Dro
p
o
u
t
f
o
r
fin
g
e
rsp
e
ll
in
g
re
c
o
g
n
it
io
n
of
Ch
in
e
se
sig
n
lan
g
u
a
g
e
,
”
M
u
lt
i
me
d
ia
T
o
o
ls
a
n
d
Ap
p
li
c
a
ti
o
n
s
,
v
o
l
.
7
9
,
n
o
.
2
1
,
p
p
.
1
5
6
9
7
-
1
5
7
1
5
,
2
0
2
0
,
d
o
i
:
1
0
.
1
0
0
7
/s1
1
0
4
2
-
0
1
9
-
0
8
3
4
5
-
y
.
[2
2
]
Q.
Zh
e
n
g
,
M.
Ya
n
g
,
J.
Ya
n
g
,
Q.
Zh
a
n
g
,
a
n
d
X.
Z
h
a
n
g
,
"
Im
p
ro
v
e
m
e
n
t
of
G
e
n
e
ra
li
z
a
ti
o
n
Ab
i
li
ty
of
De
e
p
CNN
v
ia
Im
p
li
c
it
Re
g
u
lariz
a
ti
o
n
in
Two
-
S
tag
e
Train
i
n
g
P
r
o
c
e
ss
,
"
in
I
EE
E
Acc
e
ss
,
v
o
l.
6,
p
p
.
1
5
8
4
4
-
1
5
8
6
9
,
2
0
1
8
,
d
o
i:
1
0
.
1
1
0
9
/ACCES
S
.
2
0
1
8
.
2
8
1
0
8
4
9
.
[2
3
]
X.
Xia
o
,
M.
Ya
n
,
S.
Ba
so
d
i,
C.
Ji
,
a
n
d
Y.
P
a
n
,
"
Eff
icie
n
t
h
y
p
e
r
p
a
ra
m
e
ter
o
p
ti
m
iza
ti
o
n
in
d
e
e
p
le
a
rn
in
g
u
si
n
g
a
v
a
riab
le
len
g
th
g
e
n
e
ti
c
a
l
g
o
rit
h
m
,
"
a
rXiv
p
re
p
ri
n
t
a
rXiv
:
2
0
0
6
.
1
2
7
0
3
,
2
0
2
0
.
[2
4
]
I.
Nu
r
h
a
id
a
,
V.
A
y
u
m
i,
D.
F
it
ria
n
a
h
,
R.
A.
M.
Zen
,
H.
No
p
risso
n
,
a
n
d
H.
Wei
,
“
Im
p
lem
e
n
tatio
n
of
d
e
e
p
n
e
u
ra
l
n
e
two
rk
s
(DN
N)
with
b
a
tch
n
o
r
m
a
li
z
a
ti
o
n
fo
r
b
a
ti
k
p
a
tt
e
rn
re
c
o
g
n
it
i
o
n
,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
r
n
a
l
of
El
e
c
trica
l
a
n
d
Co
mp
u
ter
E
n
g
in
e
e
rin
g
(IJ
ECE
)
,
v
o
l.
1
0
,
no
.
2,
p.
2
0
4
5
,
a
v
r
.
2
0
2
0
,
d
o
i:
1
0
.
1
1
5
9
1
/
ij
e
c
e
.
v
1
0
i2
.
p
p
2
0
4
5
-
2
0
5
3
.
[
2
5
]
P.
D
i
l
e
e
p
,
D.
D
a
s
a
n
d
P.
K.
B
o
r
a
,
"
D
e
n
s
e
L
a
y
e
r
D
r
o
p
o
u
t
B
a
s
e
d
C
N
N
A
r
c
h
i
t
e
c
t
u
r
e
f
o
r
A
u
t
o
m
a
t
i
c
M
o
d
u
l
a
t
i
o
n
C
l
a
s
s
if
i
c
a
t
i
o
n
,
"
2020
N
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
on
C
o
m
m
u
n
i
c
a
t
i
o
n
s
(NCC)
,
2
0
2
0
,
pp.
1
-
5,
d
o
i
:
1
0
.
1
1
0
9
/
N
C
C
4
8
6
4
3
.
2
0
2
0
.
9
0
5
5
9
8
9
.
[2
6
]
N.
S
riv
a
sta
v
a
,
G.
Hin
t
o
n
,
A.
Kri
z
h
e
v
sk
y
,
I.
S
u
tsk
e
v
e
r
,
a
n
d
R.
S
a
lak
h
u
td
i
n
o
v
,
"
Dr
o
p
o
u
t
:
a
sim
p
le
wa
y
to
p
re
v
e
n
t
n
e
u
ra
l
n
e
tw
o
rk
s
fr
o
m
o
v
e
rfit
ti
n
g
,
"
T
h
e
J
o
u
rn
a
l
o
f
M
a
c
h
in
e
L
e
a
rn
i
n
g
Res
e
a
rc
h
,
v
o
l.
1
5
,
n
o
.
1,
p
p
.
1
9
2
9
-
1
9
5
8
,
2
0
1
4
.
[2
7
]
P.
L
u
,
B.
S
o
n
g
,
a
n
d
L.
X
u
,
"
H
u
m
a
n
fa
c
e
re
c
o
g
n
it
i
o
n
b
a
se
d
on
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
a
n
d
a
u
g
m
e
n
te
d
d
a
tas
e
t,
"
S
y
ste
ms
S
c
ien
c
e
&
Co
n
tr
o
l
E
n
g
i
n
e
e
rin
g
,
v
o
l
.
9,
n
o
.
s
u
p
2
,
p
p
.
2
9
‑
3
7
,
M
a
y
2
0
2
1
,
d
o
i:
1
0
.
1
0
8
0
/2
1
6
4
2
5
8
3
.
2
0
2
0
.
1
8
3
6
5
2
6
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci,
Vo
l.
24
,
No
.
2
,
No
v
em
b
er
2
0
2
1
:
1
0
0
9
-
1
0
1
6
1016
[2
8
]
Ò.
Lo
re
n
te,
I.
Riera
a
n
d
A.
Ra
n
a
,
"
Im
a
g
e
Clas
sifica
ti
o
n
wi
th
Clas
sic
a
n
d
De
e
p
Lea
rn
i
n
g
Tec
h
n
i
q
u
e
s,"
Co
mp
u
ter
Vi
sio
n
a
n
d
P
a
tt
e
rn
Rec
o
g
n
it
i
o
n
,
2
0
2
1
,
a
rXiv
p
re
p
ri
n
t
a
rXiv
:2
1
0
5
.
0
4
8
9
5
,
2
0
2
1
.
[2
9
]
X.
Li
a
n
g
a
n
d
J.
Xu
,
"
Bias
e
d
Re
LU
n
e
u
ra
l
n
e
two
r
k
s,"
Ne
u
ro
c
o
mp
u
ti
n
g
,
v
o
l
.
4
2
3
,
p
p
.
7
1
‑7
9
,
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j.
n
e
u
c
o
m
.
2
0
2
0
.
0
9
.
0
5
0
.
[3
0
]
S.
-
C.
Kim
,
P.
Ra
y
,
a
n
d
S.
R.
S
a
lk
u
ti
,
“
Isla
n
d
i
n
g
d
e
tec
ti
o
n
in
a
d
istri
b
u
ti
o
n
n
e
two
r
k
wit
h
d
istri
b
u
ted
g
e
n
e
ra
to
rs
u
sin
g
sig
n
a
l
p
r
o
c
e
ss
in
g
tec
h
n
iq
u
e
s,”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
of
Po
we
r
El
e
c
tro
n
ics
a
n
d
Dr
ive
S
y
ste
ms
(IJ
PE
DS
)
,
v
o
l.
1
1
,
no
.
4,
p
p
.
2
0
9
9
-
2
1
0
6
,
2
0
2
0
,
d
o
i:
1
0
.
1
1
5
9
1
/
ij
p
e
d
s
.
v
1
1
.
i
4
.
p
p
2
0
9
9
-
2
1
0
6
.
[3
1
]
A.
Rá
c
z
,
D.
Ba
ju
sz
,
a
n
d
K.
H
é
b
e
rg
e
r,
"
Eff
e
c
t
of
Da
tas
e
t
S
iz
e
a
n
d
Trai
n
/T
e
st
S
p
li
t
Ra
ti
o
s
in
QSAR/QS
P
R
M
u
lt
icla
ss
Clas
sifica
ti
o
n
,
"
M
o
lec
u
les
,
v
o
l.
2
6
,
n
o
.
4,
p.
1
1
1
1
,
2
0
2
1
,
d
o
i
:
1
0
.
3
3
9
0
/mo
lec
u
les
2
6
0
4
1
1
1
1
.
[3
2
]
C.
M
o
o
d
ley
,
B.
S
e
p
h
to
n
,
V.
Ro
d
ríg
u
e
z
-
F
a
jard
o
,
a
n
d
A.
F
o
rb
e
s,
"
De
e
p
lea
rn
in
g
e
a
rly
sto
p
p
i
n
g
f
o
r
non
-
d
e
g
e
n
e
ra
te
g
h
o
st
ima
g
in
g
,
"
S
c
ien
ti
fi
c
Rep
o
rt
s
,
v
o
l
.
1
1
,
n
o
.
1,
p.
8
5
6
1
,
Dé
c
.
2
0
2
1
,
d
o
i:
1
0
.
1
0
3
8
/s
4
1
5
9
8
-
0
2
1
-
8
8
1
9
7
-
5.
[3
3
]
S.
S
h
re
sth
a
,
A.
Alsa
d
o
o
n
,
P.
W.
C.
P
ra
sa
d
,
I.
S
e
h
e
r
,
a
n
d
O.
H.
Al
sa
d
o
o
n
,
“A
n
o
v
e
l
s
o
lu
t
io
n
of
u
sin
g
d
e
e
p
lea
rn
i
n
g
fo
r
p
r
o
sta
te
c
a
n
c
e
r
se
g
m
e
n
tati
o
n
:
e
n
h
a
n
c
e
d
b
a
tch
n
o
rm
a
li
z
a
ti
o
n
,
”
M
u
lt
ime
d
T
o
o
ls
A
p
p
l
,
v
o
l.
8
0
,
no
.
1
4
,
pp.
2
1
2
9
3
‑2
1
3
1
3
,
Ju
n
y
2
0
2
1
,
d
o
i:
1
0
.
1
0
0
7
/s1
1
0
4
2
-
0
2
1
-
1
0
7
7
9
-
2.
[3
4
]
R.
L.
Ku
m
a
r,
J.
Ka
k
a
rla,
B.
V.
Isu
n
u
ri
,
a
n
d
M.
S
i
n
g
h
,
“
M
u
lt
i
-
c
las
s
b
ra
in
tu
m
o
r
c
las
sifica
ti
o
n
u
sin
g
re
si
d
u
a
l
n
e
two
rk
a
n
d
g
lo
b
a
l
a
v
e
ra
g
e
p
o
o
li
n
g
,
”
M
u
l
ti
me
d
T
o
o
ls
A
p
p
l
,
v
o
l.
8
0
,
no
.
9,
p
p
.
1
3
4
2
9
‑
1
3
4
3
8
,
a
v
r.
2
0
2
1
,
d
o
i:
1
0
.
1
0
0
7
/s1
1
0
4
2
-
0
2
0
-
1
0
3
3
5
-
4.
B
I
O
G
RAP
H
I
E
S
OF
AUTH
O
RS
Br
a
h
im
J
a
b
ir
wa
s
b
o
r
n
in
Az
il
a
l,
M
o
r
o
c
c
o
in
M
a
y
1,
1
9
9
0
.
He
re
c
e
iv
e
d
h
is
M
a
ste
r
d
e
g
re
e
in
2
0
1
5
in
c
o
m
p
u
ter
e
n
g
in
e
e
rin
g
a
n
d
sy
ste
m
s
at
t
h
e
S
u
lt
a
n
M
o
u
la
y
S
l
ima
n
e
Un
i
v
e
rsity
in
Be
n
i
M
e
ll
a
l,
M
o
r
o
c
c
o
.
Cu
r
re
n
tl
y
,
He
is
a
P
h
.
D.
stu
d
e
n
t
in
th
e
F
a
c
u
lt
y
of
S
c
ien
c
e
s
a
n
d
Tec
h
n
ics
of
t
h
e
sa
m
e
Un
iv
e
rsity
a
n
d
he
is
w
o
rk
i
n
g
as
a
tea
c
h
e
r
of
c
o
m
p
u
ter
sc
ien
c
e
in
th
e
re
g
i
o
n
a
l
c
e
n
ters
of
e
d
u
c
a
ti
o
n
a
n
d
train
i
n
g
p
ro
fe
ss
io
n
s
in
Be
n
i
M
e
ll
a
l,
M
o
ro
c
c
o
.
His
re
se
a
r
c
h
in
tere
sts
a
re
Dig
it
a
l
Ag
ricu
lt
u
re
,
De
e
p
lea
rn
i
n
g
,
S
trate
g
ic
An
a
ly
t
ics
a
n
d
I
n
fo
rm
a
ti
o
n
S
y
ste
m
s.
No
u
r
e
d
d
in
e
Fa
li
h
wa
s
b
o
r
n
in
Ra
b
a
t,
M
o
r
o
c
c
o
in
1
9
7
7
.
He
re
c
e
iv
e
d
h
is
P
h
D
on
C
o
m
p
u
ter
S
c
ien
c
e
s
fro
m
F
a
c
u
lt
y
of
S
c
ien
c
e
s
a
n
d
Tec
h
n
o
lo
g
ies
of
M
o
h
a
m
m
e
d
ia,
M
o
r
o
c
c
o
in
2
0
1
3
.
He
is
an
a
ss
o
c
iate
p
r
o
fe
ss
o
r
in
P
o
ly
d
is
c
ip
li
n
a
ry
F
a
c
u
l
ty
of
S
u
lt
a
n
M
o
u
l
a
y
S
l
ima
n
e
Un
i
v
e
rsity
at
Be
n
i
M
e
ll
a
l,
M
o
ro
c
c
o
si
n
c
e
2
0
1
4
.
He
h
a
s
18
y
e
a
rs
of
p
ro
fe
ss
io
n
a
l
e
x
p
e
rien
c
e
in
se
v
e
ra
l
re
n
o
w
n
e
d
c
o
m
p
a
n
ies
.
His
re
se
a
rc
h
in
tere
sts
a
re
In
fo
rm
a
ti
o
n
S
y
ste
m
G
o
v
e
rn
a
n
c
e
,
Bu
sin
e
ss
I
n
telli
g
e
n
c
e
,
Big
Da
ta
An
a
ly
t
ics
a
n
d
Di
g
it
a
l
A
g
ricu
lt
u
re
.
Evaluation Warning : The document was created with Spire.PDF for Python.